New approaches to transient detection and signal segmentation

It is of great interest to the sonar community to have automatic signal detection and segmentation procedures. In underwater passive acoustic systems one is often faced with a signal with unknown characteristics (i.e. its distribution is not completely specified), which results in a difficult detection problem, since the hypothesis test is naturally composite with structure open to challenge. Segmenting observations into different sections, mathematically modeled as detection of multiple changes, has also received extensive research attention. However, the optimum implementation is computationally formidable in practice. Accordingly this research addresses two topics: one is detection, especially of transients (short-duration signals) with partially or totally unknown characteristics; the other is segmentation of the time series. ^ The first topic covers three issues. The first issue involves robust detection of transients having unknown structure and strength. A number of improvements to Nuttall\u27s power-law detector are provided by taking advantages of tendencies toward contiguity of practical signals. The resulting detectors offer exceptional performance and are extremely easy to implement. The second issue involves detection of transients that are of unknown strength and length but with temporal contiguity. Page\u27s test is given central attention; however, a Page procedure tuned to a “short-and-loud” signal is ill-suited to a “long-but-quiet” signal. An easy adaptive alternative with time-varying thresholds is proposed and several examples are investigated to illustrate its performance improvement. The third issue involves detection of long, weak signals that are narrowband but of unknown frequency structure. An ad hoc scheme based on a multiresolution decomposition in the frequency domain is developed. Its computational load is light, and the performance is remarkably good. Generalizations are given to CFAR operation, and to the detection of multi-band signals. ^ The second topic focuses on segmentation of the time series. Two efficient and fast, alternatives are proposed to avoid a joint high-dimensional estimation of all unknowns: one is based on Gibbs sampling, and the other is based on detecting each new segment sequentially. Examples are shown for segmenting white Gaussian data with piecewise constant variances. We also extend the schemes to implement joint segmentation and classification using class-specific features.